Detecting and Localizing Electricity Pilferage Using Deep Neural Networks in a Low Voltage Distribution Network with both Smart and Traditional Meters

Electricity pilferage remains a persistent issue in the Philippines, where certain consumers resort to illegal means of utilizing power. Various studies have been conducted to devise methods for detecting the presence and location of pilferage in residential networks. Commonly used methods for detection are state estimation-based which utilize voltage and current values at grid nodes for estimation. In this study, a set of supervised deep neural networks (DNNs) were modeled and trained to estimate power values to aid in detection and localization of pilferage in a low voltage distribution network which have both smart and traditional meters, and no check meters. Errors between the readings from the meters and estimates from the DNNs were utilized to generate energy error graphs. The visualization of these errors was used to determine the presence of pilferage and possible localization. Three types of pilferage were simulated: meter bypass, diverted load, and direct tapping. It was observed that the power estimations for some houses were inaccurate. Despite this, the energy errors between the estimates and the actual readings were still helpful in the detection and localization of pilferage.